LSTM Autoencoders for EEG correction and an analysis of its latent space
Electroencephalography (EEG) is a non-invasive method used to record brain electrical activity. EEG signals are usually subject to the appearance of artifacts, which can come from a variety of sources (from ocular movements to interferences of the subject’s heartbeat), decreasing the signal to noise ratio of the EEG signal. Approaches based on Independent Component Analysis (ICA) were previously posed, showing remarkable results in artifact removal. However, these methods require human supervision. Thus, here we present a fully automated pipeline that tackles both artifact detection and real-time EEG correction. Our approach is based on a neural network within an autoencoding structure, with a core use of LSTM (long short-term memory) layers. The model was tested on electrooculographic (EOG) and electromyographic (EMG) artifacts and outperformed the main state-of-art deep learning models. In fact, it showed an extremely noticeable improvement in the case of EMG artifacts in comparison to the former approaches. Furthermore, as a side effect, the denoising autoencoder inherently leads to a representation of EEG signals into a lower dimensional space. Hence, our work also includes an exploratory analysis of the latent space learned, using linear and bilinear interpolation to furtherly understand the interactions between EEG signals in a lower dimensionality.